The Future of Robotics: Why Physical AI Matters More Than Human Appearance
Image Source: IM Imagery/Stock.adobe.com
By Poornima Apte for Mouser
Published September 25, 2026
Morgan Stanley estimates that the humanoid robot market could surpass US$5 trillion by 2050.[1] Given the scale of that projection, it is worth asking a fundamental engineering question: Must a robot take human form to do human work?
Our fascination with the humanoid robot likely traces to its pop culture roots, both in movies and literature. After all, the many different droids from Star Wars have a lingering hold on our collective imagination, as do the robot-like creatures in the Transformers franchise.
In January 2026, Boston Dynamics unveiled its newest version of a humanoid robot, Atlas, boasting 56 degrees of freedom, fully rotational joints, a reach extending to 2.28m (7.5ft), and the strength to lift up to 49.9kg (110lb).[2]
Such advances notwithstanding, the robotics industry is rife with experts both excited and skeptical about the value of humanoids and whether they’re ready for widespread adoption.
Human Abilities Matter More Than Human Appearance
The human body is an obvious blueprint for robots working in spaces built for people. But replicating that blueprint introduces engineering complexity, such as balancing on two legs and coordinating dozens of joints and actuators. This complexity can increase costs related to hardware, maintenance, and ongoing engineering, so the use of humanoid robots will need to demonstrate a significant return on investment (ROI). For now, the ROI is uncertain because not every robotics problem requires a machine that looks like a human.
The biggest draw of a general-purpose humanoid robot is that a single robot on a single platform can perform a variety of tasks, much like a human does. The fact that it looks like a human is the feature that draws the most attention, although it’s not the one that company executives likely care about.
Instead, engineers can incorporate human abilities, such as hand dexterity for object manipulation, vision to detect objects, and intelligence to process and advance goals, and transpose them to any form factor.
Next, the engineers need to determine what form the robot will take. Should it have mix-and-match human traits instead of creating a single do-it-all machine? Such fundamental questions are shaping the field of robotics, especially as physical AI enables robots to become more autonomous and respond to their environment in real time instead of simply following preprogrammed instructions.[3]
The Innovations Driving More Capable Robots
For robots to perform like humans, they need a brain (i.e., intelligence), a body (i.e., hardware), and a learning loop that learns from its mistakes and successes. Over the past decade, the industry has seen systematic advances on all these fronts.
Hardware improvements in actuators and motors have led to greater power density and finer control. Energy efficiencies in these circuits have led to improved thermal management, and a variety of end-effector advances have enabled robotic arms and fingers to manipulate objects ranging from simple rectangular boxes to packages with complex geometries. Improvements in cameras, including thermal imaging systems, combined with machine vision algorithms, have significantly enhanced robotic perception capabilities.
Physical AI plays a more significant role in the software and learning loop. Robotics foundation models are trained on broad datasets, enabling engineers to adapt them across multiple robotic tasks and platforms. One type, a vision-language-action (VLA) model, processes visual observations and language instructions to generate action representations or control commands for a robot.
Selecting the Right Human Features for the Job
While humanoid robot technology has made significant improvements in proprioception, or the ability to know where one’s limbs are in space, it might not provide the dexterity for the same set of fingers to handle both large packages and small razor blades.
The other challenge is the brittleness of the underlying models used to incorporate AI into robots. One of the key debates in robotics today is whether VLA models can generalize reliably across tasks, environments, and robot deployments, making them robust enough to serve as the foundation for general-purpose robots.[4] Part of the reason for the debate is the lack of real-world use cases for training robotics VLAs, even though robotics researchers have worked around this by creating synthetic training data.
Given that robotics engineers can start from a blank slate, the more cautious approach for the moment might be to select specific human traits that would be the most useful and use those to create modular robots.
First, engineers could consider human vision, including tasks such as object identification and depth perception. Arms are also necessary to move goods. However, the human hand’s dexterity is extremely challenging to replicate on a humanoid robot. Instead of one robot hand that can handle everything, a robot manipulation system can be modular, switching out end effectors depending on the task, from handling fragile objects to moving large bricks.
For robots to work alongside humans or in settings designed for humans, they may need to reach high platforms or shelves. Such actions do not necessarily require tall humanoid robots; instead, a robot could feature a long “neck” that can extend as needed. Additionally, cameras near the wrist or end effector, rather than on a “head,” can enable machine vision.
The Mobility Question: Legs vs. Wheels
One human trait that doesn’t always require a human form is locomotion. Walking is useful when navigating uneven terrain, but it’s also much more complex than wheels.
Wheeled robots rely on simple drivetrain mechanisms that support the vehicle’s weight, except in rare instances when the center of gravity might shift abruptly. On the other hand, the human musculoskeletal system requires joints and multiple degrees of freedom. Replicating these in a robot introduces actuators, bearings, motors, wiring, and electronic components, all of which add complexity and mass. The heavier the final product, the more energy (i.e., battery power) the robot will need, which leads to decreased utility. These issues, plus complex cable routing, additional thermal management, and highly dynamic and unpredictable loads for legged robots, make the wheeled equivalent appear more suitable for many industrial applications.
For example, Apptronik’s Apollo 2 robot supports both bipedal and wheeled configurations, with the latter being useful in high-throughput applications such as warehouses.[5] Wheeled robots are also useful inside hospitals and restaurants for equipment and order delivery.
Specialized Robots May Be Better for Some Applications
In the near term, given the human traits that can be easily mixed and matched to create robots, it might make sense to develop a fleet of specialty robots tailored for specific tasks. For example, a company may opt to use a robotic vacuum for cleaning rather than a general-purpose humanoid robot that can perform many tasks, albeit very slowly or not very well.
Such a task-based approach will likely lead to more robust and less brittle machines that can continuously improve their training models from a variety of use cases in the field. Locus Robotics, for example, uses its fleet of thousands of wheeled warehouse robots to improve adaptive awareness and warehouse automation.[6]
Using engineering to meet specific robot needs in specific situations can deliver a fleet of specialty robots in various form factors, including humanoids.
Wheeled robots include the TIAGo from PAL Robotics, which has a wheeled base that supports a height-adjustable torso, head, cameras, and optional manipulators.[7] Robots designed for healthcare settings are deliberately friendly-looking, with wheels and articulated arms. One such robot, Moxi from Diligent Robotics, is best suited for tasks like transporting medications or laboratory samples, freeing staff to focus on direct patient interactions.[8]
Legged robots include Boston Dynamics’s four-legged “canine” robot, Spot, designed especially for situations that are dangerous for humans.[7] The company also offers a bipedal humanoid robot, Atlas.
No matter what shape and form the robots take, the key takeaway is that humanoid robots should be neither dismissed altogether nor the sole focus of automation’s future.
A Future Filled with More Than Humanoid Robots
As attention-grabbing as humanoid robots are, the future is more likely to feature a variety of robots tailored to specific settings. Humanoid robots will not disappear, but the attention they are garnering today is inflated beyond near-term potential use cases.
Form typically follows function in industry. By borrowing exactly the human functions it needs to deliver measurable ROI, a fleet of specialty robots can win over executives. Advances in physical AI might also reset expectations in the near future, leading to a line of different kinds of general-purpose robots.
Whether the robots are specialty or general-purpose, there is no engineering reason for them to look like humans or to be limited to any one shape. At the 2026 Humanoid Summit in London, Nicolas Heess of DeepMind predicted that we should expect an assortment of shapes and sizes.[8] That is likely a very real outcome in an industry that is advancing rapidly.[9]
Sources
[1]https://www.morganstanley.com/insights/articles/humanoid-robot-market-5-trillion-by-2050
[2]https://bostondynamics.com/blog/boston-dynamics-unveils-new-atlas-robot-to-revolutionize-industry/
[3]https://www.ibm.com/think/topics/physical-ai
[4]https://arxiv.org/abs/2412.14058
[5]https://apptronik.com/apollo/apollo-2
[6]https://locusrobotics.com/locusone/warehouse-agentic-ai
[7]https://pal-robotics.com/robot/tiago/
[8]https://www.diligentrobots.com/moxi
[9]https://www.automate.org/robotics/blogs/how-human-does-a-humanoid-robot-need-to-look